How to Write Instructions for an AI DM Setter: Short Rules, Real Examples, No Contradictions

Good instructions for an AI DM setter are short, split into clear sections and written in plain, precise sentences. Each rule says exactly what to do, comes with a few example messages, and does not contradict any other rule. You write them yourself, change two or three things at a time, and test every change three times before you move on. AI is useful for checking your instructions, not for writing them.

Below is how I apply this when I tune instructions for clients. In the video above I do it live on real conversations.

See how the setter behaves before you write anything

Before you add a single rule, run at least one full test conversation with no instructions at all, all the way to a booked call. A good AI setter already knows how to open a conversation, ask questions and propose a time. You do not want to teach it everything from scratch. You want to see where it already works and note the two or three things you would change.

Then add those changes one at a time. Keep your old setting script or offer document in a separate doc and move pieces over only when a test conversation shows you need them.

When the setter does not do what you wanted, treat it as a gap in your instructions, not a fault in the system. In my experience, when a client tells us the setter “isn’t listening”, we usually find vague or contradictory rules in their instructions. AI is very good at guessing. When a rule leaves room, it fills the gap with the most likely answer from its training data, which is often not yours.

Short beats long: what breaks at 30,000 to 50,000 characters

Instructions of 5,000 to 6,000 characters are pleasant to work with. You can hold them in your head, scroll straight to the section you need and change it.

At 30,000, 40,000 or 50,000 characters, three things go wrong:

• The AI skips rules. The longer the instructions, the higher the chance that one of them gets ignored. That holds for practically every model.
• You lose track. Style is described in three places, the opening in two. You no longer remember where anything is, so you cannot fix it.
• Contradictions creep in. Someone adds a new rule at the bottom, forgets a rule higher up, and the two fight each other. I call that the Frankenstein.

Split the instructions into sections

Give every part of the conversation its own heading and keep each rule in its place. A simple # before the name is enough:

1. Style
2. Start of the conversation
3. Problem discovery
4. Qualification
5. Booking the call
6. Objections
7. Exceptions
8. Follow-ups

Work from the top down. The start of the conversation matters most, because a weak opening means most leads never reach your polished booking section. Get the opening right first, then move down.

Make every rule precise

Write as if you were explaining the task to a smart 12-year-old. “Be natural, be professional and don’t ask deep questions too early” gives them a rough idea, but not what to actually do.

A few examples of the difference:

• Message length. “Keep it short” is too vague. “If the lead writes 10 words, you can write at most 1.5 times as many, so 15 words” is something the model can follow. It is not perfect, because models are not great at counting, but it is far closer than “short”.
• Readiness. “Always check that the person is ready to act” means nothing to the model. Spell it out: they have 4 hours a week to train, and they are ready to commit to working on their goal for at least two months. Then add an example question.
• Conditions. Use simple logic: if X, then Y, except A and B.
• The words the system uses. When I want the setter to use what it sees on the lead’s profile, I write “profile analysis” literally, because that is what the profile summary is called inside the system. It raises the chance it will actually read it.

Give example messages in every section

Examples are gold. I aim for at least five specific rules per stage of the conversation, each with an example, and at least 12 example messages across the whole set. Put examples in quotes. For example, in the problem discovery section of a fitness offer:

Once you know how many kilos they want to lose or gain, ask what is behind it. Example: “Was there a specific moment when you decided …?”

The three dots are a placeholder. The setter fills them in from the conversation, so the template stays personal.

Give several examples per step. The setter picks the one that fits or blends them into its own reply.

Suggestion or command: decide which one you are writing

A plain rule is treated as a suggestion. In one test I set a word limit and the setter went slightly over it. That is fine for most rules: they are guidance that keeps the conversation natural.

When something must always happen, you strengthen it with words like “always”, “critical rule” or “never move on until”. When I wrote “ask about goals by sending exactly one of these examples, word for word, and change nothing”, the setter sent the example verbatim. You can steer stages the same way: “Never mark someone as disqualified until you know whether they want to lose weight or build muscle.”

Use strong words sparingly. When half your rules say “always” and “must”, the setter has to do several things at once and starts behaving oddly.

The same logic applies to scripts. I rarely give the setter a rigid script of message one, two and three. It only works when every follower arrives with the same intent, for example when all your ads and content ask people to message you for a meal plan. Most conversations do not go in a straight line. Among leads on Setor accounts who wrote first and booked a call within 30 days, the conversation took a median of 24 messages, 14 of them from the lead. A fixed sequence does not survive that.

Test every change three times in a row

After each change, run the scenario that checks it three times. If you get the reply you want three times in a row, move on. At sensitive points, or where something kept breaking, test five or ten times.

Play a different lead each time:

1. An engaged, friendly lead who follows the process.
2. A skeptical one who asks a lot of questions or changes the subject.
3. A hostile one: “I don’t believe in any of this.”

In one live session I told the setter to react with a heart when someone says no. It did so with a polite no. With a passive-aggressive lead it sent nothing, which was the smarter move. You only learn that by testing several characters.

Two more habits help. First, add a rule and check that the behavior appears, then remove it and check that it disappears. That proves the rule is doing the work. Second, take real conversations from your inbox into the test chat instead of only inventing your own. Real leads write differently from how you imagine them.

The first lessons of this method are in Simple sentences, zero contradictions: how AI understands your rules and Before you change the prompt, read real conversations with leads.

Small rules that save real conversations

These three came out of one live session on a client’s instructions:

• No repeated phrases. The setter used the same expression twice in one message. The fix: “Never use the same phrase twice. If you used a term in an earlier message, skip it or rephrase.”
• One short link. Instagram does not like many links that send people off the platform, and sending hundreds of them a month can get link sending blocked. Pin the page in a Story and send a short link to it, with one simple tracking tag instead of a long one.
• Not everything in one message. When a lead asks “details please”, the setter should not unload the whole method. Answer briefly and keep asking questions.

Cover the exceptions

Add a section for situations the setter cannot know how to handle. For example:

• When you don’t know the answer, say you will gladly come back to it, then keep the conversation going.
• When the person is already a client, ask them to use the client channel.
• When the question is about payments, refunds, terms or course access, say you look after the DMs and point them to your support email.

Only you know where your clients should go.

Use AI to diagnose your instructions, not to write them

Ask ChatGPT or Gemini to write your instructions and you get the average of everything said online about sales. It knows nothing about your offer or how your setter reads instructions. You also skip building the skill of steering your own setter. I write the instructions by hand and use AI for three checks, always with extended thinking switched on (why ready-made AI prompts are averaged rules):

The code block trick separates quotes of your rules from the AI’s comments, so you can see at a glance what to fix. Without the word “really”, the diagnostic AI gets overzealous and flags everything (every slip-up traces back to a vague sentence).

The contradiction check finds things you would not spot yourself. In one test set I planted three: “always reply in one message” against “split replies into two or three”; “no emojis, stay serious” against “praise the lead and send a flexed biceps emoji”; and “never propose a consultation until the person confirms they want it” against a four-message booking sequence that proposes the consultation in message two and asks for the go-ahead in message four. The AI caught all three, including the last one, which is a timing error the setter cannot resolve.

Last step: ask the AI to put your instructions in a logical order, fix spelling and formatting, and not change a single word where it is not absolutely necessary. Read the result and only then publish it to live conversations (final review before going live).

FAQ

How long should AI setter instructions be? As short as they can be while covering every stage. Around 5,000 to 6,000 characters is easy to maintain. At 30,000 to 50,000 characters the AI skips more rules and contradictions become hard to find.

Should I use ChatGPT to write my setter prompt? No. Use it to check your prompt: find vague rules, find contradictions and tidy the format. The rules themselves should come from you and from your real conversations.

How do I test a change to the instructions? Run the scenario that checks the change three times in a row, as an engaged lead, a skeptical one and a hostile one. Change only two or three things at a time.

Why does my AI setter ignore a rule? Usually because the rule is vague, conflicts with another rule, or is written as a suggestion when you meant a command. Make it measurable, add an example and strengthen it only if it truly must always happen.

Where the numbers come from

• 5,000 versus 30,000 to 50,000 characters, three tests per change, at least five rules per stage and 12 examples, 1.5 times the lead’s word count: my own practice from tuning setters for clients, not a study.
• 24 messages, 14 from the lead: Setor database, read on November 2, 2026. Leads who wrote first and booked a call within 30 days, 131 people, median.

Want to try this on your own DMs? Start now and write your first instructions in the test chat. If you are new to the idea, read what an AI setter is first, or see how the AI appointment setter books calls. The first message matters most, so the post on the Instagram DM opener is a good next step.